“At frontier AI labs, compute resources are typically allocated with 30% for R&D, 30% for large-scale pre-training, and 30% for inference.”